关于Oracle中各个命中率的计算以及相关的调优 1)Library Cache的命中率: .计算公式:Library Cache Hit Ratio = sum(pinhits) / sum(pins) SQL>SELECT SUM(pinhits)/sum(pins) FROM V$LIBRARYCACHE; 通常在98%以上,否则,需要要考虑加大共享池,绑定变量,修改cursor_sharing等参数。
2)计算共享池内存使用率: SQL>SELECT (1 - ROUND(BYTES / (&TSP_IN_M * 1024 * 1024), 2)) * 100 || '%' FROM V$SGASTAT WHERE NAME = 'free memory' AND POOL = 'shared pool'; 其中: &TSP_IN_M是你的总的共享池的SIZE(M)
共享池内存使用率,应该稳定在75%-90%间,太小浪费内存,太大则内存不足。 查询空闲的共享池内存: SQL>SELECT * FROM V$SGASTAT WHERE NAME = 'free memory' AND POOL = 'shared pool';
3)db buffer cache命中率: 计算公式:Hit ratio = 1 - [physical reads/(block gets + consistent gets)] SQL>SELECT NAME, PHYSICAL_READS, DB_BLOCK_GETS, CONSISTENT_GETS, 1 - (PHYSICAL_READS / (DB_BLOCK_GETS + CONSISTENT_GETS)) "Hit Ratio" FROM V$BUFFER_POOL_STATISTICS WHERE NAME='DEFAULT'; 通常应在90%以上,否则,需要调整,加大DB_CACHE_SIZE
另外一种计算命中率的方法(摘自ORACLE官方文档<<数据库性能优化>>): 命中率的计算公式为: Hit Ratio = 1 - ((physical reads - physical reads direct - physical reads direct (lob)) / (db block gets + consistent gets - physical reads direct - physical reads direct (lob)) 分别代入上一查询中的结果值,就得出了Buffer cache的命中率 SQL>SELECT NAME, VALUE FROM V$SYSSTAT WHERE NAME IN('session logical reads', 'physical reads', 'physical reads direct', 'physical reads direct (lob)', 'db block gets', 'consistent gets');
4)数据缓冲区命中率: SQL> select value from v$sysstat where name ='physical reads'; SQL> select value from v$sysstat where name ='physical reads direct'; SQL> select value from v$sysstat where name ='physical reads direct (lob)'; SQL> select value from v$sysstat where name ='consistent gets'; SQL> select value from v$sysstat where name = 'db block gets'; 这里命中率的计算应该是 令 x = physical reads direct + physical reads direct (lob) 命中率 =100 - ( physical reads - x) / (consistent gets + db block gets - x)*100 通常如果发现命中率低于90%,则应该调整应用可可以考虑是否增大数据缓冲区
5)共享池的命中率: SQL> select sum(pinhits-reloads)/sum(pins)*100 "hit radio" from v$librarycache; 假如共享池的命中率低于95%,就要考虑调整应用(通常是没使用bind var )或者增加内存
6)计算在内存中排序的比率: SQL>SELECT * FROM v$sysstat t WHERE NAME='sorts (memory)';—查询内存排序数 SQL>SELECT * FROM v$sysstat t WHERE NAME='sorts (disk)';—查询磁盘排序数 --caculate sort in memory ratio SQL>SELECT round(&sort_in_memory/(&sort_in_memory+&sort_in_disk),4)*100||'%' FROM dual; 此比率越大越好,太小整要考虑调整,加大PGA
7)PGA的命中率: 计算公式:BP x 100 / (BP + EBP) BP: bytes processed EBP: extra bytes read/written SQL>SELECT * FROM V$PGASTAT WHERE NAME='cache hit percentage'; 或者从OEM的图形界面中查看 我们可以查看一个视图以获取Oracle的建议值: SQL>SELECT round(PGA_TARGET_FOR_ESTIMATE/1024/1024) target_mb, ESTD_PGA_CACHE_HIT_PERCENTAGE cache_hit_perc, ESTD_OVERALLOC_COUNT FROM V$PGA_TARGET_ADVICE; The output of this query might look like the following:
TARGET_MB CACHE_HIT_PERC ESTD_OVERALLOC_COUNT ---------- -------------- -------------------- 63 23 367 125 24 30 250 30 3 375 39 0 500 58 0 600 59 0 700 59 0 800 60 0 900 60 0 在此例中:PGA至少要分配375M 我个人认为PGA命中率不应该低于50% 以下的SQL统计sql语句执行在三种模式的次数: optimal memory size, one-pass memory size, multi-pass memory size: SQL>SELECT name profile, cnt, decode(total, 0, 0, round(cnt*100/total,4)) percentage FROM (SELECT name, value cnt, (sum(value) over ()) total FROM V$SYSSTAT WHERE name like 'workarea exec%');
8)共享区字典缓存区命中率 计算公式:SUM(gets - getmisses - usage -fixed) / SUM(gets) 命中率应大于0.85 SQL>select sum(gets-getmisses-usage-fixed)/sum(gets) from v$rowcache;
9)数据高速缓存区命中率 计算公式:1-(physical reads / (db block gets + consistent gets)) 命中率应大于0.90最好 SQL>select name,value from v$sysstat where name in ('physical reads','db block gets','consistent gets');
10)共享区库缓存区命中率 计算公式:SUM(pins - reloads) / SUM(pins) 命中率应大于0.99 SQL>select sum(pins-reloads)/sum(pins) from v$librarycache; 11)检测回滚段的争用 SUM(waits)值应小于SUM(gets)值的1% SQL>select sum(gets),sum(waits),sum(waits)/sum(gets) from v$rollstat;
12)检测回滚段收缩次数 SQL>select name,shrinks from v$rollstat, v$rollname where v$rollstat.usn = v$rollname.usn; -----------------------------------------------------------------------------
几个常用的检查语句 1. 查找排序最多的SQL: SQL>SELECT HASH_VALUE, SQL_TEXT, SORTS, EXECUTIONS FROM V$SQLAREA ORDER BY SORTS DESC;
2.查找磁盘读写最多的SQL:
SQL>SELECT * FROM (SELECT sql_text,disk_reads "total disk" , executions "total exec",disk_reads/executions "disk/exec" FROM v$sql WHERE executions>0 and is_obsolete='N' ORDER BY 4 desc) WHERE ROWNUM<11 ;
3.查找工作量最大的SQL(实际上也是按磁盘读写来排序的): SQL>select substr(to_char(s.pct, '99.00'), 2) || '%' load,s.executions executes,p.sql_text from(select address,disk_reads,executions,pct,rank() over (order by disk_reads desc) ranking from (select address,disk_reads,executions,100 * ratio_to_report(disk_reads) over () pct from sys.v_$sql where command_type != 47) where disk_reads > 50 * executions) s,sys.v_$sqltext p where s.ranking <= 5 and p.address = s.address order by 1, s.address, p.piece;
4. 用下列SQL工具找出低效SQL: SQL>select executions,disk_reads,buffer_gets,round((buffer_gets-disk_reads)/buffer_gets,2) Hit_radio,round(disk_reads/executions,2) reads_per_run,sql_text From v$sqlarea Where executions>0 and buffer_gets >0 and (buffer_gets-disk_reads)/buffer_gets<0.8 Order by 4 desc;
5、根据sid查看对应连接正在运行的sql SQL>select /*+ push_subq */command_type,sql_text,sharable_mem,persistent_mem,runtime_mem,sorts,version_count,loaded_versions,open_versions,users_opening,executions,users_executing,loads,first_load_time,invalidations,parse_calls,disk_reads,buffer_gets,rows_processed,sysdate start_time,sysdate finish_time,’>’||address sql_address,’N’status From v$sqlarea Where address=(select sql_address from v$session where sid=&sid);
***************Oracle 缓冲区命中率低的分析及解决办法****************** 首先确定下面的查询结果: 1,缓冲区命中率的查询(是否低于90%): select round((1 - sum(decode(name,'physical reads',value,0)) / (sum(decode(name,'db block gets',value,0)) + sum(decode(name,'consistent gets',value,0))) ),4) *100 || '%' chitrati from v$sysstat; 2,使用率的查询(有无free状态的数据快.): select count(*), status from v$bh group by status ; 3,相关等待事件的查询(是否有相关等待事件) select event,total_waits from v$system_event where event in ('free buffer waits'); 4,当前大小(是否已经很大) select value/1024/1024 cache_size from v$parameter where name='db_cache_size' 5,top等待事件分析(Db file scatered read的比率是否大) select event ,total_waits,suml from (select event,total_waits,round(total_waits/sumt*100,2)||'%' suml from (select event,total_waits from v$system_event ), (select sum(total_waits) sumt from v$system_event) order by total_waits desc) where rownum<6 and event not like 'rdbms%' and event not like 'pmon%' and event not like 'SQL*Net%' and event not like 'smon%'; 6,db_cache_advice建议值(9i后的新特性,可以根据他更好的调整cache_size) select block_size,size_for_estimate,size_factor,estd_physical_reads from v$db_cache_advice; 说明分析: 缓冲区命中率(低于90的命中率就算比较低的). 没有free不一定说明需要增加,还要结合当前cache_size的大小(我们是否还可以再增大,是否有需要增加硬件,增加开销), 空闲缓冲区等待说明进程找不到空闲缓冲区,并通过写出灰缓冲区,来加速数据库写入器生成空闲缓冲区,当DBWn将块写入磁盘后,灰数据缓冲区将被释放,以便重新使用.产生这种原因主要是: 1,DBWn可能跟不上写入灰缓冲区:i/0系统较慢,尽量将文件均匀的分布于所有设备, 2,缓冲区过小或过大。 3,可以增加db_writer_processes数量。 4,可能有很大的一个事物,或者连续的大事物 我们需要长期观察这个事件是否长期存在并数值一直在增大,如果一直在增大,则说明需要增大db_cache大小.或优化sql. 数据分散读等待,通常表现存在着与全表扫描相关的等待,逻辑读时,在内存中进行的全表扫描一般是零散地,而并非连续的被分散到缓冲区的各个部分,可能有索引丢失,或被仰制索引的存在。该等待时间在数据库会话等待多块io读取结束的时候产生,并把指定的块数离散的分布在数据缓冲区。这意味这全表扫描过多,或者io不足或争用, 存在这个事件,多数都是问题的,这说明大量的全部扫描而未采用索引. db_cache_advice对我们调整db_cache_size大小有一定的帮助,但这只是一个参考,不一定很精确。 通过上面6种情况的综合分析,判断是否需要增加大cache_size. 或者把常用的(小)表放到keep区。 但多数的时候做这些不会解决质的问题, 而真正的问题主要是对sql语句的优化(如:是否存在大量的全表扫描等) 索引是在不需要改变程序的情况下,对数据库性能,sql语句提高的最实用的方法. 我在生产中遇到过类似的问题,200M的cache_size,命中率很低21%,但通过对sql语句的优化(添加索引,避免全表扫描),命中率增加到96%,程序运行时间由原来的2小时减少到不到10分钟. 这就提到了怎么定位高消耗的sql问题.全表扫描的问题,在这里不做细致的解说,这里只说明方法,我会在相关的章节专门介绍怎么使用这些工具 1,sql_trace跟踪session.用tkprof 分别输出磁盘读,逻辑读,运行时间长的sql进行优化.这些高消耗的sql一般都伴随着全表扫描. 2,statspack分析.在系统繁忙时期进行时间点的统计分析,产看TOP事件是否有Db file scatered read.并查看TOP sql语句是否存在问题等. 注:电脑学习网首发。 还要说一句:当然在硬件允许的情况下,尽量增大db_cache_size 减少磁盘读,但并不是越大越好,一定要根据自己的库数据量的程度来调节,因为大的db_cache_size同样会增大数据库管理的开销,当然可能开销并不会明显的影响数据库的性能,硬件价格也越来越低,这就需要我们具体问题具体分析了,在我看来物尽其用就最好了,尽量不要浪费,找到问题的本质。调优是一件很艺术的事。 ***********************Oracle数据库缓冲区命中率*****************
1、查看Oracle数据库缓冲区命中率
select a.value + b.value "logical_reads", c.value "phys_reads", round(100 * ((a.value+b.value)-c.value) / (a.value+b.value)) "BUFFER HIT RATIO" from v$sysstat a, v$sysstat b, v$sysstat c where a.statistic# = 40 and b.statistic# = 41 and c.statistic# = 42; 2、Tags: oracle
数据库缓冲区命中率: sql>select value from v$sysstat where name ='physical reads'; value 3714179 sql>select value from v$sysstat where name ='physical reads direct'; value 0 sql>select value from v$sysstat where name ='physical reads direct(lob)'; value 0 sql>select value from v$sysstat where name ='consistent gets'; value 856309623 sql>select value from v$sysstat where name ='db block gets'; value 19847790 这里命中率的计算应该是 令x=physical reads direct + physical reads direct(lob) 命中率=100-(physical reads -x)/(consistent gets +db block gets -x)*100 通常如果发现命中率低于90%,则应该调整应用可以考虑是否增大数据加
共享池的命中率 sql> select sum(pinhits)/sum(pins)*100 "hit radio" from v$librarycache; 如果共享池的命中率低于95%就要考虑调整应用(通常是没应用bind var)或者增加内存。
关于排序部分 sql> select name,value from v$sysstat where name like '%sort%'; 如果我们发现sorts(disk)/(sorts(memory)+sorts(disk))的比例过高,则通常意味着sort_area_size部分内存教较小,可考虑调整相应的参数。
关于log_buffer sql>select name,value from v$sysstat where name in ('redo entries','redo buffer allocation retries');
假如redo buffer allocation retries/redo entries的比例超过1%我们就可以考虑增加log_buffer. |